Bespoke kernels

A bridge trained on your distribution.

We run our frontier diffusion model over your features and return a lightweight mixture model -- weights and all. All run on our distributed GPU system.

What ships back

01

A fitted GMM bridge

Drift weights for a Gaussian-mixture Schrödinger bridge fit to your target samples — not a generic checkpoint with your data bolted on.

02

A CUDA kernel

Compiled, benchmarked, and deliberately small. Sub-millisecond evaluation on a single device, with no framework runtime required at inference time.

03

The training run

Shard logs, convergence traces, and a holdout evaluation — so you can see how the fit behaved rather than taking the weights on faith.

Your infrastructure stays empty

Training runs on our bridge cluster — parallel shards across GPU workers, aggregated into one set of drift weights. You never provision a training box, and you never wait behind your own queue.

Small enough to embed

A mixture bridge is a fraction of the size of a diffusion model at comparable sample quality on structured data. That's the whole point: it fits where a general-purpose generator won't.

Plans

Free to configure. Pay when it trains.

Sign up and you can stage a dataset, set up a fit and read the SDK without paying anything. The Developer plan is what puts batch jobs on our compute — bridge training, flow matching and hosted inference against the kernels that come back.

A fit is one kernel trained against one dataset. Volume, annual terms and training inside your own VPC are scoped as an engagement.

Send us your samples.

Upload a target distribution and we'll tell you what the fit will look like before you commit to it.

Upload training data